I've been using Claude Opus 5.5 since it came out yesterday on September 22, and for the first time since early this year I don't feel like I'm managing the model. It answers the question, it's quick, and it doesn't try to explain everything when unnecessary. Opus 4.7 through 5.0 never quite got there for me and it doesn't really feel great to use. This feels a lot closer to Opus 4.6, the version a lot of us still talk about the way people talk about a favorite old car. That's a feeling, so here's what's behind it.
Opus 5 wrote long and felt frustrating. It narrated, it verified things nobody asked it to verify, and it treated a one-line fix like a project, which is most of what we wrote about in August. Anthropic's own announcement for 5.5 says the new model "used a third of the tokens Opus 5 did, and its answers were 40% less verbose without losing accuracy." It also says output is more than 30 percent faster. You notice both within about five minutes.
It writes plainly finally. Every ran its own tests on launch day, and Katie Parrott's review says "Opus 5.5 produces the most readable prose we've seen from an Anthropic or OpenAI model." The headline on that review is that it's pulling their Codex converts back to Claude, and most of the early reaction I've seen elsewhere has been praise.
And usage seems to go further as I'm on the 20x subscription and it feels like the weekly usage % is going up slower doing the same work. Anthropic also raised five-hour limits with this release and added a rate limit reset you can save and spend when you want it. None of that undoes the 17 percent weekly cut from September 14, but it's the first limits news in a while that points in the right direction.
Which effort setting to use
Opus 5.5 is the first Opus to ship with medium as its default effort. Every earlier model that supports effort defaults to high. Anthropic's pitch is that "at default effort (medium), Opus 5.5 beats GPT-6 Astra at max effort for about a fifth of the cost per task."
Medium or high is where I'd stay and is probably the best bang for the buck, and I said the same about Opus 5 in July. Anthropic's effort docs describe medium as the "balanced approach with moderate token savings," meant for agentic tasks that need a balance of speed, cost, and performance, and high as the level for complex reasoning and difficult coding. My split is simple. Medium for everyday work, and high when I hit something medium keeps getting wrong.
Low is for simple tasks where speed and cost matter most, and the docs name subagents as the example. Xhigh is meant for long-running agentic and coding work, over 30 minutes, with token budgets in the millions. Max has no limits on token spending, and on a subscription that mostly means hitting your cap sooner.
Please don't nerf it
Every time a model feels this good, somebody starts a thread a week or two later asking whether it got worse and it got nerfed. Sometimes it's just the honeymoon wearing off. Sometimes something really did change. In a September 2025 postmortem, Anthropic said that "between August and early September, three infrastructure bugs intermittently degraded Claude's response quality," and also that "we never reduce model quality due to demand, time of day, or server load."
Regardless of what they say, I sometimes feel like the models get nerf and degrade over time. That's why I hope there is no nerf.
Additionally, for the Haiku fans who are still on 4.5 (!), Sonnet 5.5 and Haiku 5.5 are due in the coming weeks, according to Anthropic.
Sources
Anthropic: Introducing Claude Opus 5.5 - The September 22, 2026 release, the token and verbosity reduction versus Opus 5, the 30 percent faster output, the medium default effort and GPT-6 Astra cost comparison, the five-hour limit increase and savable rate limit reset, and the coming Sonnet 5.5 and Haiku 5.5 releases.
Every: Vibe Check, Opus 5.5 is pulling our Codex converts back to Claude - Katie Parrott's September 22, 2026 review, including the readability finding, the training-schedule failure, the note that it still buries the point, unconstrained token use, and the comparison with Fable 5.1.
Anthropic docs: Effort - Medium as the Opus 5.5 default versus high on other models, the descriptions of each effort level, and the guidance to run a fresh effort sweep instead of carrying settings over.
Anthropic: A postmortem of three recent issues - The September 17, 2025 postmortem on three infrastructure bugs that degraded response quality, and the statement that Anthropic never reduces model quality due to demand, time of day, or server load.
